부산대학교 도서관

This study proposes a series of models for reconstructing 3D point clouds from a single 2D image to obtain the best network model. Four and six improved models are proposed for the encoder and decoder of 3D-LMNet and 3D-PSRNet, respectively, which combine various modules of such encoders and decoders and analyze the relationship of each parameter to the network layer. Optimal allocation parameters are proposed, and four training types are presented for the encoder and decoder to obtain the best model. The model adds a fifth convolution layer to the 3D-PSRNet coding layer. This layer has 512 layers. The convolution size is set to 5 × 5 and the stride is 2. The proposed model does not require professional hardware equipment and cumbersome manual procedures.